What the documents look like
Multi-warehouse allocation happens at the order routing stage and determines which warehouse ships an order. The working materials come from WMS, OMS, ERP inventory and platform inventory interfaces, and the typical roles involved are supply chain planning, WMS administrators and e-commerce operations. The daily document surface consists of allocation rules, inventory synchronization tables, transfer orders and shortage alerts. Errors in this area lead to oversell refunds, higher cross-region freight or delayed shipment. These materials update daily or hourly, so parsing and chunking must keep conditions and constraints inside the same material traceable back to the original text.
Allocation rules usually carry warehouse priority and applicability conditions. Inventory synchronization tables carry sellable quantity and warehouse identifiers. Transfer orders carry source and destination warehouses plus quantities. Shortage alerts carry thresholds and product scope. If chunking separates a condition from the content it constrains, a retrieved fragment will lack its subject or its effective range.
In FastGPT, the knowledge base consists of library, collection and data. The smallest search unit is the library; a collection only classifies data and does not affect search results. One data item can map to multiple vectors. During retrieval, vector IDs return to the dataset.datas table in MongoDB to merge original data belonging to the same group, and the score takes the highest value. The vector retriever uses the vector field and HNSW index in PostgresSQL, while MongoDB stores other data. These structures mean chunking must preserve original identity and version so that a fragment can be traced back for verification.
Constraints at the document chunking step
The chunking step must keep the subject, status and validity range of each material intact. A transfer order marked as pending review is not a completed transfer. An ERP same-day snapshot basis is not a synchronization action. A similarity threshold filters low-relevance results and does not perform permission or entity isolation. Retrieval is not authorization.
Hypothetical paper exercise. The source material contains three sentences. "East China warehouse ships with priority from 2025-06-01 to 2025-06-30; sellable inventory follows the ERP same-day snapshot." "Transfer order DB-0612: 120 units transferred out from South China warehouse to East China warehouse, status pending review." "Shortage alert threshold: a certain SKU in the East China warehouse triggers when sellable quantity falls below 30 units." The draft sentence to be judged is: "The South China warehouse has completed the transfer of 120 units to the East China warehouse, and East China sellable inventory has been synchronized." This sentence treats a pending-review status as completed and treats a snapshot basis as a synchronization action. Both claims lack support in the source material.
Configuration and implementation checks
The table below covers this page's sample. Values are hypothetical starting points for this exercise, used for comparison against the sample text, and are not official defaults.
| Configuration or check item | Hypothetical starting point | Reason and comparison object |
|---|---|---|
| Search mode | Hybrid search | Sample contains identifier DB-0612 and natural-language conditions; compare against full-text only and semantic only |
| Reference limit | Measured in tokens, starting point 2000 | Official control is by tokens; compare against whether the three sample sentences are fully covered |
| Minimum relevance | Starting point 0.4 | Takes effect only under semantic search or rerank; compare against the draft sentence relevance |
| Question optimization | Not enabled in this exercise | Sample question is already complete; compare against the added latency when enabled |
| Result rerank | Not enabled in this exercise | Requires a ReRank model enabled per settings; compare against scores when enabled |
| Post-update review | Manual check of training and index status | Upload success does not equal retrieval update; compare against the source material version |
Three common errors
- Treating a pending-review transfer order as a completed transfer. Input is the draft sentence and the transfer order source text. Compare the status field "pending review" against "completed". The observable result is that the draft sentence lacks a basis for completion.
- Writing an ERP same-day snapshot basis as synchronized. Input is the draft sentence and the inventory synchronization table source text. Compare "follows the ERP same-day snapshot" against "has been synchronized". The observable result is that the synchronization action has no source support.
- Using a similarity setting to exclude unrelated business. Input is the minimum relevance value and the sample query. Compare the documented statement that this value only filters low-relevance results. The observable result is that this value does not carry permission or entity isolation.
How to verify the setup
- Input the three source sentences and the draft sentence, then under hybrid search compare whether recalled fragments contain subject, status and validity range. If a fragment is missing "pending review" or "2025-06-30", the completeness check is not satisfied.
- Input identifier DB-0612 as a full-text search and compare whether the transfer order fragment is hit. If it is not hit, record this as pending verification and do not conclude the fragment does not exist.
- Input an empty result [] as a paper check and compare whether the downstream design handles it as an empty reference. Do not claim that a search for a fictional rule necessarily returns []. Libraries across different vector models cannot be associated together; first check identifiers and exclude non-matching ones, then record the actual observation.